More details on Fable 5’s cyber safeguards and our jailbreak framework
Anthropic details Fable 5's cybersecurity safeguards and releases jailbreak testing framework for red-teaming.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Anthropic details Fable 5's cybersecurity safeguards and releases jailbreak testing framework for red-teaming.
Just for kicks, I took a look at Jersey Mike's IPO documents. Surely a sandwich shop would have no need to mention AI. But low-and-behold.
Simon Willison releases llm-coding-agent 0.1a0, a Python agent framework for LLM-based code generation with file and command execution tools.
Meta has quietly launched Pocket, an experimental AI app that lets users generate and share interactive mini games using text prompts.
The news comes about a week after OpenAI announced its own custom AI chip in a partnership with Broadcom.
Simon Willison uses DSPy to evaluate and optimize SQL system prompts in Datasette Agent, a tool enabling agents to execute read-only database queries.
Introduces Iterative VibeCoding, a benchmark for studying prompt-injection and code-injection attacks in autonomous AI agents shipping iterative PRs to persistent codebases.
LACUNA benchmark evaluates whether LLM unlearning methods truly erase PII from model parameters or merely obfuscate it, testing against resurfacing attacks.
Program-as-Weights paradigm compiles fuzzy functions from natural-language specs into locally-executable 4B neural artifacts, with 10M-example FuzzyBench dataset.
Risk-controlled real-time safety monitor for LLM outputs at deployment time, thresholding verifier signals to flag unsafe generations with mathematical guarantees.
ReContext: training-free inference method using recursive evidence replay to improve long-context reasoning by exploiting model-internal relevance signals.
Dual-channel debate framework reveals that LLM agents develop latent objectives and shift public utterances based on social structure, audience, and role without explicit prompting.
DramaSR-532K benchmark for speaker recognition in TV dramas with 532K annotated dialogue lines; applies reasoning LLMs to improve multimodal attribution.
DemoPSD addresses privileged information leakage in on-policy LLM self-distillation by using disagreement-modulated training to prevent answer-dependent shortcuts.
Systematic comparison of matrix-structured optimizers (Muon, SOAP) vs. Adam for training machine-learning interatomic potentials on NequIP and Allegro models.
CNeVA: controllable simulated agents with interpretable behavior latents for realistic traffic simulation, enabling edge-case testing without real-world risk.
CLIP vision encoders vulnerable to typographic attacks; training-free concept localization proposed for robustness in safety-critical applications.
G-RRM neuro-symbolic framework guides symbolic solvers with symbol-equivariant recurrent reasoning models for constraint satisfaction problems.
VRRL reinforcement learning framework improves vision-language model self-reflection by grounding corrections in visual inputs for out-of-distribution robustness.
EADP framework addresses token pruning in VLMs via entropy-aware structured compression, reducing inference cost while preserving dense instruction comprehension.
Audio feature analysis from LibriVox correlates narrator acoustics with audiobook consumption; limited direct AI relevance.
TestEvo-Bench executable benchmark evaluates test-code co-evolution for automated testing agents; mined from real repositories.
Polymarket study finds human-AI forecasting outcomes driven by human capital, not model performance; majority of users either defer or rubber-stamp.
Task-Agnostic Pretraining decouples motor learning from semantic alignment in Vision-Language-Action models to reduce expert demonstration bottleneck.
Scaling laws investigation suggests LLM social simulation fidelity may not improve with compute scaling; proposes dedicated research attention.
OrbitQuant data-agnostic quantization for diffusion transformers; uses randomized permuted block-Hadamard rotation for stable timestep-invariant inference.
Neuron-aware data selection improves LLM self-distillation in annotation-free post-training without domain-specific labels.
Philosophical critique of LLMs as cultural measurement tools, arguing models constitute rather than neutrally measure cultural phenomena.
Theoretical analysis of distributed self-supervised learning robustness under non-IID data; Masked Image Modeling shows inherent robustness.
Quantum algorithm study on stabilizer state testing with limited coherent memory; theoretical contribution tangential to LLM frontier.